Autonomous Quadrotor Control with Reinforcement Learning

Based on the same principles as a single-rotor helicopter, a quadrotor is a flying vehicle that is propelled by four horizontal blades surrounding a central chassis. Because of this vehicle’s symmetry and propulsion mechanism, a quadrotor is capable of simultaneously moving and steering by simple modulation of motor speeds [1]. This stability and relative simplicity makes quadrotors ideal for research in the application of control theory and artificial intelligence to aerial robotics [3]. Most prior work using quadrotors has applied low-level, manually-tuned control algorithms to complete specific tasks. This paper proposes an alternate approach for controlling a quadrotor through the application of continuous stateaction space reinforcement learning algorithms by making use of the Parrot AR.Drone’s rich suite of on-board sensors and the localization accuracy of the Vicon motion tracking system. With such high quality state information a reinforcement learning algorithm should be capable of quickly learning a policy that maps the quadrotor’s physical state to the low level velocity parameters that are used to control a quadrotors’s four motors. Once learning is complete, this policy will encode the information necessary to repeatably and accurately perform the desired high-level action without ever requiring a programmer to manually split the action into smaller components.